-
- Private work.
- Read through some Bedrock (from old colleague): https://caylent.com/blog/amazon-bedrock-everything-you-need-to-know
- Tightened the comment settings on this blog (all spam).
- Aquarium.
- Tomato clown bonded to the old bubbletip (evicting the 2 old perculas). Sebae clown bonded to the new bubbletip. Old ocellaris still homeless.
- Sally lightfoot was apex for quite some time. Now the engineering gobies are large enough and have started nipping at the crabs.
- Deep RL.
- OpenAI’s docs: https://spinningup.openai.com/
- Install
spinningup and libopenmpi-dev and mujoco-py.
- OpenMPI = Message Passage Interface (for HPC): https://www.open-mpi.org/
- MuJoCo = Multi Joint dynamics with Contact. It’s a physics engine. https://mujoco.readthedocs.io/. OpenAI maintains a python lib for it: https://github.com/openai/mujoco-py
- Neural network libs: pytorch, tensorflow.
- OpenCV = Computer Vision. Python bindings for this (wheel takes a bit of time to build fyi).
- Can run multiple algorithms, policies, sims, plot outputs, more.
- Some good educational overviews on there too: policies for what actions to take in which states, cost/reward functions, Bellman equations.
- Algos: Vanilla Policy Gradient (VPG), Trust Region Policy Optimization (TRPO), Proximal Policy Optimization (PPO), Deep Deterministic Policy Gradient (DDPG), Twin Delayed DDPG (TD3), Soft Actor-Critic (SAC).
- Switched trays, first tightening.
- Colab is Google’s jupyter implementation: https://colab.research.google.com/. Python in the browser.
- TensorFlow.
- Ran 5-10 notebooks to play with some of the functionality.
- Watched https://www.youtube.com/playlist?list=PLQY2H8rRoyvwWuPiWnuTDBHe7I0fMSsfO
- Fitting data, pulling different models, training in other ways.
- Comes with a bunch of existing datasets to train against (
keras.datasets). Computer vision example: 100,000 images of cats and dogs, and a classification of each as cat or dog. Then the trained model can see new images and predict the classification.
- Specify
loss (like mean squared, how to measure inaccuracy) and optimizer (how to choose the next guess).
- You can use convoluted filters for feature extraction. Basically just many different layers, which one best produces output.
- This was probably my fav notebook from the examples: https://www.tensorflow.org/tutorials/keras/classification
- Crypto.
- Submitted withdrawal request for gOhm from tokemak was week. Today noon was cycle rollover, so was able to complete the withdrawal.
- Swapped directly for gOHM -> USDC on uniswap (+ one extra transaction to approve gOHM). Could have “unstaked” on olympus to convert gOHM -> OHM, but that’s another unnecessary tx.
- Then transferred USDC from metamask to coinbase, converted to USD, withdrew to bank, and transferred to TD for equities.
- Overall, I expected to get crushed by olympus; I hadn’t checked it in about 1yr and crypto has fallen considerably. I thought this token would be ~10% of my basis. It was 94%!
- The metamask dapp has some cool portfolio tools: https://portfolio.metamask.io/
- The new (well, I haven’t used in a long time) coinbase advanced trade interfaces is much better than the old pro.coinbase: https://www.coinbase.com/advanced-trade
- Supercontest.
- Banner/lines/picks.
- Westgate posted MNF football with the date a week off (the monday prior), so my app flagged the lines as old and didn’t commit them.
- https://gitlab.com/bmahlstedt/supercontest/-/issues/215
- Emergency alert system ran a test on all phones at 2:18pm ET.
- Went to Jazba in EV, Junoon’s new spinoff.
- Overdrafted BoA by accident.
- Emptied hydroponics. Will do full clean tomorrow, and replant soon after.
- Updated vscode.
